Scalable SCADA-driven failure prediction for offshore wind turbines using autoencoder-based NBM and fleet-median filtering

Offshore wind turbines are crucial for sustainable energy production but face significant challenges in operational reliability and maintenance costs. In particular, the scalability and practicality of failure detection systems are a key challenge in large-scale wind farms. This paper presents a scalable, comprehensive approach to failure prediction based on the normal behavior modeling (NBM) framework that integrates three components: a cloud-based pipeline, an undercomplete autoencoder for temperature-based anomaly detection, and a time-aware anomaly filtering method. The pipeline enables dynamic scaling and streamlined deployment across multiple wind farms. The autoencoder was trained exclusively on healthy 10 min SCADA data and produces detailed anomaly scores that serve as the input for our filtering technique. It was trained on 4 years of data from a large offshore wind farm in the Dutch–Belgian zone and achieved unhealthy–healthy (UHH) ratios of up to 1.69 and 1.21 for the generator and gearbox models, respectively. The filtering method refines the raw anomaly scores by comparing turbine signals to a windowed fleet median. By aggregating scores via sliding windows and employing robust distance metrics, the method reduces the volume of anomaly scores by up to 65 % without sacrificing predictive accuracy. This selective filtering effectively minimizes noise and non-relevant anomalies, enhancing the efficiency of maintenance analysis.